This may not going to work for sociolinguistics? It’s still a self-supervised learner. And now we can point to the original data for each word. Hmm. I take it back. THis could be cool